📍 Forthcoming in IEEE Region 10 Conference (TENCON 2026) • Bali, Indonesia

Orthographic Nativization as Rule-Based Rewrite Cascades

A linguistically grounded, priority-ordered rewrite cascade for adapting English loanwords into standardized Filipino orthography without parallel training corpora.

Erin Gabrielle Chua Zhean Robby Ganituen Justin Ethan Ching Jaztin Jacob Jimenez Nathaniel Oco

College of Computer Studies, De La Salle University, Manila, Philippines
MangoCats Research Groupzhean_robby_ganituen@dlsu.edu.ph

Research Motivation

The Orthographic Nativization Challenge

Generating standardized written forms of loanwords in low-resource recipient languages.

Abstract & Linguistic Formulation

Orthographic nativization—generating standardized written forms of loanwords in a recipient language—remains understudied computationally despite its importance for text normalization, spell-checking, and multilingual NLP. When English loanwords enter Filipino, they undergo substantial phonotactic and graphemic changes prescribed by the official Komisyon sa Wikang Filipino (KWF) guidelines (e.g., computerkompyuter, straightistreyt, scheduleiskedyul).

Unlike abstract phonological adaptation, orthographic nativization requires producing concrete character sequences obeying the target language's orthography. TagaBaybay operationalizes prescriptive linguistic rules and phonological descriptions into a deterministic four-stage rewrite cascade, eliminating the need for expensive parallel training corpora.

Linguistic Grounding

Directly formalizes the codified prescriptive manual by Virgilio Almario (KWF 2014) alongside phonological descriptions of Philippine English and Filipino.

Zero Parallel Corpora

Deployed with zero training data requirements. Deterministic execution produces zero hallucinations, sub-millisecond throughput, and fully interpretable rewrite steps.

5.14% Character Error Rate

Evaluated on 2,319 gold-standard pairs, outperforming both regular expressions (20.38% CER) and zero-shot LLMs with identical rules (15.54% CER).

System Architecture

Four-Stage Rewrite Pipeline

An English loanword passes through tokenization, context-sensitive cascades, phonetic lookup, and normalization.

TagaBaybay Orthographic Nativization Pipeline Diagram

Figure 1: Orthographic nativization pipeline. An English loanword is tokenized into multigraphs, passed through priority-ordered rewrite rules and G2P phonetic resolution, followed by repair and cleanup.

1
Graphemic Tokenization
Greedy longest-match multigraph tokenization identifying tetragraphs (augh, eigh), trigraphs (tch, ore), and digraphs (ph, ch, sh).
2
Context-Sensitive Rules
Priority-ordered rewrite cascade testing neighbor contexts (e.g. c before front vowels e/i/ys, otherwise k; geminate collapse ccks).
3
Phonetic Resolution (G2P)
Consults upstream grapheme-to-phoneme lookup to resolve ambiguous vowels (such as unstressed schwa /ə/ or glide expansions /juː/).
4
Output Normalization
Final repairs: sC cluster prosthesis (scheduleis-), glide insertion (hiatus repair), silent -e truncation, and orthographic cleanup.
Interactive Showcase

Rewrite Cascade in Action

Select an English loanword below to inspect its grapheme tokenization and priority-ordered rewrite cascade.

Choose an English loanword:
Filipino Nativized Form /kəmˈpjuː.tɚ/
kompyuter
Stage 1: Graphemic Tokens
Stages 2 & 3: Rewrite Cascade Trace Rule Application
Experimental Evaluation

Benchmark Results on 2,319 Loanwords

Evaluated against a newly constructed gold standard of English–Filipino loanword pairs under multiple vowel-equivalence conditions.

5.14%
5-Vowel CER
Strict evaluation
4.05%
3-Vowel CER
Accounting for e/i and o/u
2,319
Gold Standard Pairs
Constructed test dataset
12.59%
CER Reduction
vs. spelling-only ablation
Method / Model Paradigm 5-Vowel CER 3-Vowel CER Notes / Behavior
TagaBaybay (Proposed) Rule-Based Rewrite Cascade + G2P 5.14% 4.05% Highest accuracy; deterministic & interpretable
Spelling-Only Ablation Rules without G2P Resolution 17.73% 16.02% Fails on ambiguous vowels and schwa reductions
Large Language Model (LLM) Zero-Shot Prompting (Rules in Prompt) 15.54% 14.10% Cannot reliably execute character-level constraints
Regular Expressions Baseline Unordered Pattern Substitutions 20.38% 19.12% Suffers from rule collisions & context bleed
Overall CER across vowel equivalence conditions

Figure 2: Overall CER (%) under four vowel-equivalence conditions from the paper.

Primary Error Analysis Takeaways

  • 1.
    Unstressed Schwa Ambiguity: English schwa /ə/ lacks an exact Filipino counterpart, leading to auditory substitutions between e, i, and a.
  • 2.
    Vowel Hiatus Glide Insertion: Glides /w/ and /j/ break vowel hiatus (e.g. tiaratiyara).
  • 3.
    Upstream G2P Quality: 40 character edits traced directly to pronunciation lexicons on long pharmaceutical terms.
Research Team

Project Members

Faculty Adviser: Nathaniel Oco

We would also like to thank the following people for their contributions to an earlier prototype of the system:

    Citation

    Found Our Work Helpful?

    BibTeX
    @inproceedings{tagabaybay2026tencon,
      author    = {Chua, Erin Gabrielle and Ganituen, Zhean Robby and Ching, Justin Ethan and Jimenez, Jaztin Jacob and Oco, Nathaniel},
      title     = {Orthographic Nativization as Rule-Based Rewrite Cascades},
      booktitle = {Proceedings of the 2026 IEEE Region 10 Conference (TENCON)},
      location  = {Bali, Indonesia},
      year      = {2026},
      publisher = {IEEE}
    }
    
    @software{tagabaybay_code,
      author    = {Ganituen, Zhean Robby and Chua, Erin Gabrielle and Ching, Justin Ethan and Jimenez, Jaztin Jacob and Ang, Clive Jarel and Ang, Clarence Ivan and Campo, Roan Cedric and Oco, Nathaniel},
      title     = {{TagaBaybay: A Phonetic Nativization Algorithm for Filipino Loanwords}},
      year      = {2026},
      url       = {https://github.com/Mango-Cats/tagabaybay},
      license   = {Apache-2.0}
    }
    Communication & Inquiry

    Send us a Question

    [TagaBaybay]
    The subject starter [TagaBaybay] is automatically included.